Home Blog About Me Recommendations
- 16 Dec, 2025 *
Recently I have been experimenting with unusual methods of using LLMs. This all started when I read a tweet (currently lost somewhere in my bookmarks) about a developer’s process where he asked Claude to generate detailed implementation plans for large projects before bed. In the morning Claude would have gathered all the context and created a plan. He would somewhat follow this plan, but it was mostly a way of staying focused on the project while allowing for serendipity.
What interested me about this process was it was the first t…
Home Blog About Me Recommendations
- 16 Dec, 2025 *
Recently I have been experimenting with unusual methods of using LLMs. This all started when I read a tweet (currently lost somewhere in my bookmarks) about a developer’s process where he asked Claude to generate detailed implementation plans for large projects before bed. In the morning Claude would have gathered all the context and created a plan. He would somewhat follow this plan, but it was mostly a way of staying focused on the project while allowing for serendipity.
What interested me about this process was it was the first time when I saw AI being used to support the flow state rather than trying to avoid the deep work of programming. It could provide supporting material while you do the thinking.
That brings me onto my bizarre method. I quite like writing a small, meandering essay about a topic I am interested in without much research. It’s more of a hypothesis than an essay. Normally only 5 or 6 paragraphs of loose prose. I take this fragment and stick it into Claude and ask it to research around the topic. Whatever interesting sources are surfaced I dig into myself, reformulate my thinking, and repeat the process.
I recently tried this with an essay about how video might actually be an incredibly poor method of educating people online, this stems from the huge amount of edutainment on platforms like YouTube. I was surprised by the number of intriguing and relevant sources, especially on productive failure and the paradox that great lecturers can actually make you feel like you’re learning much more than you actually are.
In the above case, I am essentially using Claude as a librarian. It’s not doing any of the work of writing for me, but is instead letting me dip into some research I might not otherwise have found. Key-word search like Google can’t really do this since I am not always sure of the technical terms required to surface the high-quality sources. Claude is a LLM, which is very good at matching concepts over matching keywords.
The process of trying to formulate an argument for something by writing a short, private essay followed by reading actual research on the topic has proven to be a very fun, deep way to learn. I’ve found that even with ADHD I’ve been able to return to an argument over the course of many weeks and slowly develop my thinking around a topic without depending on the AI to think for me.
For anyone who enjoys writing for fun, this might be a way to learn subjects organically without surrendering your thinking to an LLM.